SIMD128 intrinsics support? #9255
Description
Feature request
What is the expected behavior?
When writing WASM (from Rust) that uses SIMD128 intrinsics, I get this error:
ERROR in ../pkg/test_wasm_bg.wasm
Module parse failed: Internal failure: parseVec could not cast the value
You may need an appropriate loader to handle this file type.
Error: Internal failure: parseVec could not cast the value
at new CompileError (/Users/nsthorat/Code/rust/test-wasm/www/node_modules/@webassemblyjs/helper-api-error/lib/index.js:40:109)
at parseVec (/Users/nsthorat/Code/rust/test-wasm/www/node_modules/@webassemblyjs/wasm-parser/lib/decoder.js:328:15)
at parseTypeSection (/Users/nsthorat/Code/rust/test-wasm/www/node_modules/@webassemblyjs/wasm-parser/lib/decoder.js:352:29)
at parseSection (/Users/nsthorat/Code/rust/test-wasm/www/node_modules/@webassemblyjs/wasm-parser/lib/decoder.js:1347:24)
at Object.decode (/Users/nsthorat/Code/rust/test-wasm/www/node_modules/@webassemblyjs/wasm-parser/lib/decoder.js:1696:25)
at decode (/Users/nsthorat/Code/rust/test-wasm/www/node_modules/@webassemblyjs/wasm-parser/lib/index.js:248:21)
I tracked it down to webassemblyjs (I don't totally understand the relationship between these projects, so maybe it needs to be filed there):
https://github.com/xtuc/webassemblyjs/blob/4ddbd20f23a320d134895829001374cfa1da84de/packages/wasm-parser/src/decoder.js#L354
What is motivation or use case for adding/changing the behavior?
I work on TensorFlow.js and we're trying to get SIMD128 support for vectorized machine learning primitives.
How should this be implemented in your opinion?
Webpack & webassemblyjs shouldn't fail.
Are you willing to work on this yourself?
yes
Coming from Rust, here is some code:
let mut a: Vec<f32> = vec![0.0; 4];
for i in 0..a.len() {
a[i] = i as f32;
}
let raw = a.as_ptr() as *const core_arch::wasm32::v128;
let lane1 = core_arch::wasm32::v128_load(raw);
let out = core_arch::wasm32::f32x4_extract_lane(lane1, 0);
The generated WASM looks like this:
call $core::slice::<impl__T_>::as_ptr::h561ee13c3c552cb4
call $core_arch::core_arch::wasm32::simd128::v128_load::ha5b99643436b0934
call $core_arch::core_arch::wasm32::simd128::f32x4_extract_lane::h2aaffb364379d64d
It's worth noting that some intrinsics work.